Toxic Comment Filter

1 Introduction

Build a model that can filter user comments based on the degree of language maliciousness:

• Preprocess the text by eliminating the set of tokens that do not make significant contribution at the semantic level.
• Transform the text corpus into sequences.
• Build a Deep Learning model including recurrent layers for a multilabel classification task.
• At prediction time, the model should return a vector containing a 1 or a 0 at each label in the dataset (toxic, severe_toxic, obscene, threat, insult, identity_hate). In this way, a non-harmful comment will be classified by a vector of only 0s [0,0,0,0,0]. In contrast, a dangerous comment will exhibit at least a 1 among the 6 labels.

2 Setup

Leveraging Quarto and RStudio, I will setup an R and Python enviroment.

2.1 Import R libraries

Import R libraries. These will be used for both the rendering of the document and data analysis. The reason is I prefer ggplot2 over matplotlib. I will also use colorblind safe palettes.

library(tidyverse, verbose = FALSE)
library(tidymodels, verbose = FALSE)
library(reticulate)
library(ggplot2)
library(plotly)
library(RColorBrewer)
library(bslib)
library(Metrics)
library(gt)

reticulate::use_virtualenv("r-tf")

2.2 Import Python packages

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import keras
import keras_nlp

from keras.backend import clear_session
from keras.models import Model, load_model
from keras.layers import TextVectorization, Input, Dense, Embedding, Dropout, GlobalAveragePooling1D, LSTM, Bidirectional, GlobalMaxPool1D, Flatten, Attention
from keras.metrics import Precision, Recall, AUC, SensitivityAtSpecificity, SpecificityAtSensitivity, F1Score


from sklearn.model_selection import train_test_split, KFold
from sklearn.metrics import multilabel_confusion_matrix, classification_report, ConfusionMatrixDisplay, precision_recall_curve, f1_score, recall_score, roc_auc_score

Create a Config class to store all the useful parameters for the model and for the project.

2.3 Class Config

I created a class with all the basic configuration of the model, to improve the readability.

class Config():
    def __init__(self):
        self.url = "https://s3.eu-west-3.amazonaws.com/profession.ai/datasets/Filter_Toxic_Comments_dataset.csv"
        self.max_tokens = 20000
        self.output_sequence_length = 911 # check the analysis done to establish this value
        self.embedding_dim = 128
        self.batch_size = 32
        self.epochs = 100
        self.temp_split = 0.3
        self.test_split = 0.5
        self.random_state = 42
        self.total_samples = 159571 # total train samples
        self.train_samples = 111699
        self.val_samples = 23936
        self.features = 'comment_text'
        self.labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
        self.new_labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate', "clean"]
        self.label_mapping = {label: i for i, label in enumerate(self.labels)}
        self.new_label_mapping = {label: i for i, label in enumerate(self.labels)}
        self.path = "/Users/simonebrazzi/R/blog/posts/toxic_comment_filter/history/f1score/"
        self.model =  self.path + "model_f1.keras"
        self.checkpoint = self.path + "checkpoint.lstm_model_f1.keras"
        self.history = self.path + "lstm_model_f1.xlsx"
        
        self.metrics = [
            Precision(name='precision'),
            Recall(name='recall'),
            AUC(name='auc', multi_label=True, num_labels=len(self.labels)),
            F1Score(name="f1", average="macro")
            
        ]
    def get_early_stopping(self):
        early_stopping = keras.callbacks.EarlyStopping(
            monitor="val_f1", # "val_recall",
            min_delta=0.2,
            patience=10,
            verbose=0,
            mode="max",
            restore_best_weights=True,
            start_from_epoch=3
        )
        return early_stopping

    def get_model_checkpoint(self, filepath):
        model_checkpoint = keras.callbacks.ModelCheckpoint(
            filepath=filepath,
            monitor="val_f1", # "val_recall",
            verbose=0,
            save_best_only=True,
            save_weights_only=False,
            mode="max",
            save_freq="epoch"
        )
        return model_checkpoint

    def find_optimal_threshold_cv(self, ytrue, yproba, metric, thresholds=np.arange(.05, .35, .05), n_splits=7):

      # instantiate KFold
      kf = KFold(n_splits=n_splits, shuffle=True, random_state=42)
      threshold_scores = []

      for threshold in thresholds:

        cv_scores = []
        for train_index, val_index in kf.split(ytrue):

          ytrue_val = ytrue[val_index]
          yproba_val = yproba[val_index]

          ypred_val = (yproba_val >= threshold).astype(int)
          score = metric(ytrue_val, ypred_val, average="macro")
          cv_scores.append(score)

        mean_score = np.mean(cv_scores)
        threshold_scores.append((threshold, mean_score))

        # Find the threshold with the highest mean score
        best_threshold, best_score = max(threshold_scores, key=lambda x: x[1])
      return best_threshold, best_score

config = Config()

3 Data

The dataset is accessible using tf.keras.utils.get_file to get the file from the url. N.B. For reproducibility purpose, I also downloaded the dataset. There was time in which the link was not available.

# df = pd.read_csv(config.path)
file = tf.keras.utils.get_file("Filter_Toxic_Comments_dataset.csv", config.url)
df = pd.read_csv(file)

library(reticulate)

py$df %>%
  tibble() %>% 
  head(5)

Table 1: First 5 elemtns

# A tibble: 5 × 8
  comment_text            toxic severe_toxic obscene threat insult identity_hate
  <chr>                   <dbl>        <dbl>   <dbl>  <dbl>  <dbl>         <dbl>
1 "Explanation\nWhy the …     0            0       0      0      0             0
2 "D'aww! He matches thi…     0            0       0      0      0             0
3 "Hey man, I'm really n…     0            0       0      0      0             0
4 "\"\nMore\nI can't mak…     0            0       0      0      0             0
5 "You, sir, are my hero…     0            0       0      0      0             0
# ℹ 1 more variable: sum_injurious <dbl>

Lets create a clean variable for EDA purpose: I want to visually see how many observation are clean vs the others labels.

df.loc[df.sum_injurious == 0, "clean"] = 1
df.loc[df.sum_injurious != 0, "clean"] = 0

3.1 EDA

First a check on the dataset to find possible missing values and imbalances.

3.1.1 Frequency

library(reticulate)
df_r <- py$df
new_labels_r <- py$config$new_labels

df_r_grouped <- df_r %>% 
  select(all_of(new_labels_r)) %>%
  pivot_longer(
    cols = all_of(new_labels_r),
    names_to = "label",
    values_to = "value"
  ) %>% 
  group_by(label) %>%
  summarise(count = sum(value)) %>% 
  mutate(freq = round(count / sum(count), 4))

df_r_grouped %>% 
  gt() %>% 
  tab_header(
    title = "Labels frequency",
    subtitle = "Absolute and relative frequency"
  ) %>% 
  fmt_number(
    columns = "count",
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = TRUE,
    use_seps = TRUE
  ) %>% 
  fmt_percent(
    columns = "freq",
    decimals = 2,
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = FALSE
  ) %>% 
  cols_align(
    align = "center",
    columns = c("count", "freq")
  ) %>% 
  cols_align(
    align = "left",
    columns = label
  ) %>% 
  cols_label(
    label = "Label",
    count = "Absolute Frequency",
    freq = "Relative frequency"
  )

Table 2: Absolute and relative labels frequency

Labels frequency
Absolute and relative frequency
Label	Absolute Frequency	Relative frequency
clean	143,346	80.33%
identity_hate	1,405	0.79%
insult	7,877	4.41%
obscene	8,449	4.73%
severe_toxic	1,595	0.89%
threat	478	0.27%
toxic	15,294	8.57%

3.1.2 Barchart

library(reticulate)
barchart <- df_r_grouped %>%
  ggplot(aes(x = reorder(label, count), y = count, fill = label)) +
  geom_col() +
  labs(
    x = "Labels",
    y = "Count"
  ) +
  # sort bars in descending order
  scale_x_discrete(limits = df_r_grouped$label[order(df_r_grouped$count, decreasing = TRUE)]) +
  scale_fill_brewer(type = "seq", palette = "RdYlBu") +
  theme_minimal()
ggplotly(barchart)

Figure 1: Imbalance in the dataset with clean variable

It is visible how much the dataset in imbalanced. This means it could be useful to check for the class weight and use this argument during the training.

It is clear that most of our text are clean. We are talking about 0.8033 of the observations which are clean. Only 0.1967 are toxic comments.

3.2 Sequence lenght definition

To convert the text in a useful input for a NN, it is necessary to use a TextVectorization layer. See the Section 4 section.

One of the method is output_sequence_length: to better define it, it is useful to analyze our text length. To simulate what the model we do, we are going to remove the punctuation and the new lines from the comments.

3.2.1 Summary

library(reticulate)
df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
    ) %>% 
  pull(text_length) %>% 
  summary() %>% 
  as.list() %>% 
  as_tibble()

Table 3: Summary of text length

# A tibble: 1 × 6
   Min. `1st Qu.` Median  Mean `3rd Qu.`  Max.
  <dbl>     <dbl>  <dbl> <dbl>     <dbl> <dbl>
1     4        91    196  378.       419  5000

3.2.2 Boxplot

library(reticulate)
boxplot <- df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
    ) %>% 
  # pull(text_length) %>% 
  ggplot(aes(y = text_length)) +
  geom_boxplot() +
  theme_minimal()
ggplotly(boxplot)

Figure 2: Text length boxplot

3.2.3 Histogram

library(reticulate)
df_ <- df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
  )

Q1 <- quantile(df_$text_length, 0.25)
Q3 <- quantile(df_$text_length, 0.75)
IQR <- Q3 - Q1
upper_fence <- as.integer(Q3 + 1.5 * IQR)

histogram <- df_ %>% 
  ggplot(aes(x = text_length)) +
  geom_histogram(bins = 50) +
  geom_vline(aes(xintercept = upper_fence), color = "red", linetype = "dashed", linewidth = 1) +
  theme_minimal() +
  xlab("Text Length") +
  ylab("Frequency") +
  xlim(0, max(df_$text_length, upper_fence))
ggplotly(histogram)

Figure 3: Text length histogram with boxplot upper fence

Considering all the above analysis, I think a good starting value for the output_sequence_length is 911, the upper fence of the boxplot. In the last plot, it is the dashed red vertical line.. Doing so, we are removing the outliers, which are a small part of our dataset.

3.3 Dataset

Now we can split the dataset in 3: train, test and validation sets. Considering there is not a function in sklearn which lets split in these 3 sets, we can do the following: - split between a train and temporary set with a 0.3 split. - split the temporary set in 2 equal sized test and val sets.

x = df[config.features].values
y = df[config.labels].values

xtrain, xtemp, ytrain, ytemp = train_test_split(
  x,
  y,
  test_size=config.temp_split, # .3
  random_state=config.random_state
  )
xtest, xval, ytest, yval = train_test_split(
  xtemp,
  ytemp,
  test_size=config.test_split, # .5
  random_state=config.random_state
  )

xtrain shape: py$xtrain.shape ytrain shape: py$ytrain.shape xtest shape: py$xtest.shape ytest shape: py$ytest.shape xval shape: py$xval.shape yval shape: py$yval.shape

The datasets are created using the tf.data.Dataset function. It creates a data input pipeline. The tf.data API makes it possible to handle large amounts of data, read from different data formats, and perform complex transformations. The tf.data.Dataset is an abstraction that represents a sequence of elements, in which each element consists of one or more components. Here each dataset is creates using from_tensor_slices. It create a tf.data.Dataset from a tuple (features, labels). .batch let us work in batches to improve performance, while .prefetch overlaps the preprocessing and model execution of a training step. While the model is executing training step s, the input pipeline is reading the data for step s+1. Check the documentation for further informations.

train_ds = (
    tf.data.Dataset
    .from_tensor_slices((xtrain, ytrain))
    .shuffle(xtrain.shape[0])
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

test_ds = (
    tf.data.Dataset
    .from_tensor_slices((xtest, ytest))
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

val_ds = (
    tf.data.Dataset
    .from_tensor_slices((xval, yval))
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

print(
  f"train_ds cardinality: {train_ds.cardinality()}\n",
  f"val_ds cardinality: {val_ds.cardinality()}\n",
  f"test_ds cardinality: {test_ds.cardinality()}\n"
  )

train_ds cardinality: 3491
 val_ds cardinality: 748
 test_ds cardinality: 748

Check the first element of the dataset to be sure that the preprocessing is done correctly.

train_ds.as_numpy_iterator().next()

(array([b"The only way that what Marc Shepherd said would be done most likely is if there's a Wikiproject for for NYC Transit, MTA, and MaBSTOA buses, which I doubt is needed or wanted.",
       b'I have added the paragraph that answers this question.',
       b'"\nthanks, I\'ll hunt some up.  (Talk) "',
       b"Put 'em put 'em you hermit crab lovin gekko photographing homun72.221.92.43  culus.",
       b', rather than a genuine concern over a particular policy violation',
       b'"\nComment Hi There, In spite all my effort, I don\xe2\x80\x99t expect any understanding or help from the other side regarding the English version of the article about Alexander Voytovych. Therefore if one insists on deleting this particular article then go ahead as it seems to me there is no room for democratic thinking on your website whatsoever. I\xe2\x80\x99d like to add a few more words as my final statement. Wikipedia says:  \xe2\x80\x9cSources may encompass published works in all forms and media, and in any language. http://en.wikipedia.org/wiki/Wikipedia:Notability. The art works and art life of the artist is the sourcefor academic press, Ukrainian media, academic institutions, as well as poets, writers and publications use his paintings & drawings to illustrate books. The artist takes a part in a busy art life in the Ukraine presenting a number of exhibitions & art projects. Having said that I see that there is no enough evidence for one????? \nTABBodnar O./\xd0\x91\xd0\xbe\xd0\xb4\xd0\xbd\xd0\xb0\xd1\x80 \xd0\x9e. \xe2\x80\x94 \xd0\x9b. \xd0\xa2\xd0\xb0\xd0\xb9\xd0\xbd\xd0\xb0\xd0\xb2\xd1\x82\xd1\x96\xd0\xbb\xd0\xb5\xd0\xbd\xd0\xbd\xd1\x8f. \xe2\x80\x94 \xd0\xa3\xd0\xb6\xd0\xb3\xd0\xbe\xd1\x80\xd0\xbe\xd0\xb4: \xd0\x9a\xd0\xb0\xd1\x80\xd0\xbf\xd0\xb0\xd1\x82\xd0\xb8, 2009. \xe2\x80\x94 \xd0\xa1. 237\xe2\x80\x94239. \xe2\x80\x94 f\xd0\xbet\xd0\xbe: \xd1\x81. 227, 270.ISBN 978-966-671-179-6 \nTABGavrosh A./\xd0\x93\xd0\xb0\xd0\xb2\xd1\x80\xd0\xbe\xd1\x88 \xd0\x9e. \xd0\xa2\xd1\x96\xd0\xbb\xd0\xbe\xd0\xbb\xd1\x83\xd1\x87\xd0\xbd\xd0\xb8\xd1\x86\xd1\x96. \xe2\x80\x94 \xd0\x9b.:\xd0\x9f\xd1\x96\xd1\x80\xd0\xb0\xd0\xbc\xd1\x96\xd0\xb4\xd0\xb0, 2006. \xe2\x80\x94 56 \xd1\x81. \xe2\x80\x94 \xd0\x86\xd0\xbb\xd1\x8e\xd1\x81\xd1\x82\xd1\x80\xd0\xb0\xd1\x86\xd1\x96\xd1\x97: \xd0\x9e.\xd0\x92\xd0\xbe\xd0\xb9\xd1\x82\xd0\xbe\xd0\xb2\xd0\xb8\xd1\x87. ISBN 966-8522-69-6 \nTABDidyk N./\xd0\x94\xd1\x96\xd0\xb4\xd0\xb8\xd0\xba \xd0\x9d. Fundamentals of composition. \xe2\x80\x94 \xd0\xa3\xd0\xb6\xd0\xb3\xd0\xbe\xd1\x80\xd0\xbe\xd0\xb4: \xd0\x9c\xd0\xb8\xd1\x81\xd1\x82\xd0\xb5\xd1\x86\xd1\x8c\xd0\xba\xd0\xb0\xd0\xbb\xd1\x96\xd0\xbd\xd1\x96\xd1\x8f, 2009. \xe2\x80\x94 \xd0\xa1.44. ISBN 978-966-8764-96-7 \nTABKosmolinska N./\xd0\x9a\xd0\xbe\xd1\x81\xd0\xbc\xd0\xbe\xd0\xbb\xd1\x96\xd0\xbd\xd1\x81\xd1\x8c\xd0\xba\xd0\xb0 \xd0\x9d. Art chat on ""Green sofa"" // FINE ART. \xe2\x80\x94 2009. \xe2\x80\x94 \xe2\x84\x96 4. \xe2\x80\x94 \xd0\xa1. 114 \xe2\x80\x94 115. \nTABChervatiuk L./\xd0\xa7\xd0\xb5\xd1\x80\xd0\xb2\xd0\xb0\xd1\x82\xd1\x8e\xd0\xba \xd0\x9b. Women\'s image in the Ukrainian modern art. \xe2\x80\x94 \xd0\x9a.:\xd0\x9d\xd0\xb0\xd0\xb2\xd1\x87\xd0\xb0\xd0\xbb\xd1\x8c\xd0\xbd\xd0\xb0\xd0\xba\xd0\xbd\xd0\xb8\xd0\xb3\xd0\xb0, 2007. \xe2\x80\x94 C. 33 \xe2\x80\x94 34.ISBN 978-966-329-110-9 *TABShumylovych B./\xd0\xa8\xd1\x83\xd0\xbc\xd0\xb8\xd0\xbb\xd0\xbe\xd0\xb2\xd0\xb8\xd1\x87 \xd0\x91. Women, paintings and allegories in high heels. // \xd0\x9e\xd0\xb1\xd1\x80\xd0\xb0\xd0\xb7\xd0\xbe\xd1\x82\xd0\xb2\xd0\xbe\xd1\x80\xd1\x87\xd0\xb5\xd0\xbc\xd0\xb8\xd1\x81\xd1\x82\xd0\xb5\xd1\x86\xd1\x82\xd0\xb2\xd0\xbe. \xe2\x80\x94 2008. \xe2\x80\x94 \xe2\x84\x96 4. \xe2\x80\x94 \xd0\xa1. 92 \xe2\x80\x94 93. \nIf one convinced that the article has to be deleted, than do not hesitate to do so, but before that please give me a straight answer to my question. What is the difference between the article about Alexander Voytovych and, for example, the Herb Roe\xe2\x80\x99s article, why the mentioned article is considered more accurate? I must highlight once again that instead of explaining what has to be changed or corrected on the page, one awards us with thenoticeboards. Regarding self-promotion  which was mentioned a few times I\xe2\x80\x99d like to make it clear  the artists is a Ukrainian and naturally all sources are Ukrainian, and if the mentioned language is unknown to one it doesn\xe2\x80\x99t meant that it allows one doubts the genuine sources. The Ukrainian and Russian versions of Wikipedia have accepted the article and honoured it, which says a lot.I had in my mind to present more Ukrainian artist, however having such an experience; I\xe2\x80\x99d rather leave this business behind thanks to you. Thanks for co-operation. From my point of view this discussion is exceeded, feel free to delete the article!\n  "',
       b'"\n\nSpeedy deletion of Seby torreblanca\n Please do not make personal attacks. Wikipedia has a strict policy against personal attacks. Attack pages and images are not tolerated by Wikipedia and are speedily deleted. Users who continue to create or repost such pages and images in violation of our biographies of living persons policy will be blocked from editing Wikipedia. Thank you. \n\nIf you think that this notice was placed here in error, you may contest the deletion by adding  to the top of the page (just below the existing speedy deletion or ""db"" tag), coupled with adding a note on the talk page explaining your position, but be aware that once tagged for speedy deletion, if the article meets the criterion it may be deleted without delay. Please do not remove the speedy deletion tag yourself, but don\'t hesitate to add information to the article that would would render it more in conformance with Wikipedia\'s policies and guidelines.    Talk  "',
       b"No defense for firing?  This was a man who was found guilty of sexual harassment by a government investigator, and then later convicted in a court of law.  If you have information that is not in the public domain, such as the secret legal opinion obtained by Lorne Calvert's government concerning the firing of Murdoch Carriere, then I would implore you to table that opinion here (or on the internet).  Otherwise, all that can be concluded is that it is sheer misconduct and bungling on the part of Calvert and his government that led to the payment of $275,000 plus additional sums relating to pension credits.\n\nTable the legal opinion if you have it, or tell your cronies in the party to convince Calvert to table the opinion under the immunity of Parliamentary Priviledge.  Otherwise, your beloved leader and party will continue to fall victim to the criticisms that have been levelled against it, and will not be able to credibly defend itself.",
       b'"\n\nU.S. support to Juarez, especially in the form of weapons, came after the end of the American Civil War. They also allowed Juarez\' government to operate for some time from American soil. That doesn\'t seem like ""little help"" to me.  e  | \xcf\x84\xce\xb1\xce\xbb\xce\xba "',
       b'even that security council resolution states that israel has put the golan under civilian law, they just dont recongnize its annexation. i agree with the marbehtorah change it to Territories under ISraeli control209.255.127.242',
       b'"Hi.  I\'m from the Willamette congregation Eugene, OR.  Sorry don\'t know any McGhee\'s in California.  I like your ""Opposition"" idea.  Cheers, Matt M. )      03:42, 16 Jun 2005 (UTC)\n\n"',
       b'Is Smith married? Gay? What?',
       b'Kosovo and Serbia\nSame edit war. Same problem. See Talk:Serbia. 204.52.215.107',
       b"They are different volcanoes.  I unfortunately haven't got a photo of Hl\xc3\xb6\xc3\xb0ufell to upload but here's a link to one: .  I can't seem to find a decent map of Iceland online, but Hl\xc3\xb6\xc3\xb0ufell is southwest of Langj\xc3\xb6kull and Her\xc3\xb0ubrei\xc3\xb0 is north of Vatnaj\xc3\xb6kull (just northeast of Askja).  I would think that Her\xc3\xb0ubrei\xc3\xb0 is considered Iceland's most famous tuya and I haven't heard any such sentiment about Hl\xc3\xb6\xc3\xb0ufell.  Neither of the names mean tuya, but I think the best translation of tuya might be B\xc3\xbarfell (which name many Icelandic mountains possess).  Best regards,",
       b'Ogstrokes and 24.239.149.9 are the same person.   24.239.149.9 was before registration, and OGstrokes is after registration.',
       b'At Bookfinder.com I found Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference\n\n Softcover, ISBN: 1558604790 Publisher: Morgan Kaufmann Pub, 1988\n\n Bookfinder',
       b'It was not an attack on someone, she is a personal friend of mine and i was doing it as a temporary gag. As soon as i was told it was going to be deleted, i deleted the contents accordingly. yes the page should have been deleted but it was not an attack on anyone.',
       b'"\nI countered your complaint about ""operation iraqi freedom"" with the fact that it is not even the article title and thus changing the link to a neutral term in no way shows bias. I removed ""just cause"" just because it is a propaganda term.  "',
       b'"\nThen I will just request you to please upload the image. It is the most appropriate image for the article. \'\'\'\'\'\'  "',
       b'TUSC token eaf3126a18fae1f097f990290c84672a \n\nI am now proud owner of a TUSC account!',
       b"Yes Bollyjeff, that is my thought, its not her debut film. If you go around web, you'll find all fancruft sites saying 2002,2003 but, you can check album details of film in allnusic, it says 2003. Music was released in2003  but film was delayed, much like Ishq In Paris whose music was released but now it seems forgettable.You got my point what I'm saying.",
       b'"\nHere\'s a link you may find helpful. Wikipedia:Mediation. Charge! "',
       b"This album is very similar to Clayman in all aspects, except perhaps it have a softer approach. So why don't leave just melodeath and hard rock",
       b'"Based on reports in the media this week, and from comments made by others who have experienced what I went through, it\'s apparently not uncommon in Wikipedia for a single person to create several different personalities, or ID\'s, and to then use them themselves, or in concert with others, to create the appearance of unrelated people making edits, challenging posters, harrassing and threatening others, and worse. So, your statment that ""..hmm, no, it was three people."", rings hollow with some of us.\n \n\n"',
       b'RfA thanks \n\nThank you very much for your support at my RfA. Regards,  (talk)',
       b'"\n\n Shouldn\'t byte range locking be mentioned for the Windows platform? \n\nFor Unix locking, it is mentioned that ""different kinds of locks may be applied to different sections (byte ranges) of a file"". This is not mentioned for the Windows platform. \n\nHowever, when I checkout the LockFile function in msdn, it seems to me that this is also possible on the windows platform, though only for ""server"" editions of the OS. I\'m not a file-locking specialist, so I\'m hesitating a bit to edit the article..."',
       b'Myspace worm\n\nDo we really need to mention the FSM Myspace worm? It seems awfully unrelated to the actual Pastafarianism concept. -',
       b'http://indianmilitarynews.files.wordpress.com/2011/05/indian-army-rape-us.jpg\nIndian kids trying to rape my user page ( 86.182.174.123',
       b'"\nGo spill your philosophies elsewhere. This isn\'t a church.  (mailbox) "',
       b"No, the reference is talking about Total Shakira's album sales in US, is not Sale el Sol WW sales. Nielsen SoundScan only counts US sales, only IFPI give WW sales.",
       b"Followup: Ten edits from this IP (to date) include seven that were quickly reverted by other editors as worthless. Five of those seven (plus one more constructive edit) relate to emergency medicine, and four of the five are bad jokes/deletable nonsense involving Dunkin' Donuts, dating back more than a year (to edits of Paramedic on 30 March 2007). The difference between running gag and repeat vandalism is subtle, but nonetheless real: It's comedy shows that have running gags, whereas freely-editable encyclopedias have repeat vandalism. If we have any more trouble with this, I believe the evidence would get us an edit ban.",
       b"Now who's all up in arms about cultural minutiae. Face it bud, you didn't do you homework. Nor did you discuss on the talk page before you slapped a tag on it. I'm sure you can find some sort of wikijustification for your poor research. Just keep to stuff you know from here on out. Otherwise you'll get your ass handed to you over and over again. 71.102.2.128"],
      dtype=object), array([[0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [1, 0, 0, 0, 0, 0]]))

And we check also the shape. We expect a feature of shape (batch, ) and a target of shape (batch, number of labels).

print(
  f"text train shape: {train_ds.as_numpy_iterator().next()[0].shape}\n",
  f" text train type: {train_ds.as_numpy_iterator().next()[0].dtype}\n",
  f"label train shape: {train_ds.as_numpy_iterator().next()[1].shape}\n",
  f"label train type: {train_ds.as_numpy_iterator().next()[1].dtype}\n"
  )

text train shape: (32,)
  text train type: object
 label train shape: (32, 6)
 label train type: int64

4 Preprocessing

Of course preprocessing! Text is not the type of input a NN can handle. The TextVectorization layer is meant to handle natural language inputs. The processing of each example contains the following steps: 1. Standardize each example (usually lowercasing + punctuation stripping) 2. Split each example into substrings (usually words) 3. Recombine substrings into tokens (usually ngrams) 4. Index tokens (associate a unique int value with each token) 5. Transform each example using this index, either into a vector of ints or a dense float vector.

For more reference, see the documentation at the following link.

text_vectorization = TextVectorization(
  max_tokens=config.max_tokens,
  standardize="lower_and_strip_punctuation",
  split="whitespace",
  output_mode="int",
  output_sequence_length=config.output_sequence_length,
  pad_to_max_tokens=True
  )

# prepare a dataset that only yields raw text inputs (no labels)
text_train_ds = train_ds.map(lambda x, y: x)
# adapt the text vectorization layer to the text data to index the dataset vocabulary
text_vectorization.adapt(text_train_ds)

This layer is set to: - max_tokens: 20000. It is common for text classification. It is the maximum size of the vocabulary for this layer. - output_sequence_length: 911. See Figure 3 for the reason why. Only valid in "int" mode. - output_mode: outputs integer indices, one integer index per split string token. When output_mode == “int”, 0 is reserved for masked locations; this reduces the vocab size to max_tokens - 2 instead of max_tokens - 1. - standardize: "lower_and_strip_punctuation". - split: on whitespace.

To preserve the original comments as text and also have a tf.data.Dataset in which the text is preprocessed by the TextVectorization function, it is possible to map it to the features of each dataset.

processed_train_ds = train_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)
processed_val_ds = val_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)
processed_test_ds = test_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)

5 Model

5.1 Definition

Define the model using the Functional API.

def get_deeper_lstm_model():
    clear_session()
    inputs = Input(shape=(None,), dtype=tf.int64, name="inputs")
    embedding = Embedding(
        input_dim=config.max_tokens,
        output_dim=config.embedding_dim,
        mask_zero=True,
        name="embedding"
    )(inputs)
    x = Bidirectional(LSTM(256, return_sequences=True, name="bilstm_1"))(embedding)
    x = Bidirectional(LSTM(128, return_sequences=True, name="bilstm_2"))(x)
    # Global average pooling
    x = GlobalAveragePooling1D()(x)
    # Add regularization
    x = Dropout(0.3)(x)
    x = Dense(64, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.01))(x)
    x = LayerNormalization()(x)
    outputs = Dense(len(config.labels), activation='sigmoid', name="outputs")(x)
    model = Model(inputs, outputs)
    model.compile(optimizer='adam', loss="binary_crossentropy", metrics=config.metrics, steps_per_execution=32)
    
    return model

lstm_model = get_deeper_lstm_model()
lstm_model.summary()

5.2 Callbacks

Finally, the model has been trained using 2 callbacks: - Early Stopping, to avoid to consume the kaggle GPU time. - Model Checkpoint, to retrieve the best model training information.

# callbacks
my_es = config.get_early_stopping()
my_mc = config.get_model_checkpoint(filepath="/checkpoint.keras")
callbacks = [my_es, my_mc]

5.3 Final preparation before fit

Considering the dataset is imbalanced, to increase the performance we need to calculate the class weight. This will be passed during the training of the model.

lab = pd.DataFrame(columns=config.labels, data=ytrain)
r = lab.sum() / len(ytrain)
class_weight = dict(zip(range(len(config.labels)), r))
df_class_weight = pd.DataFrame.from_dict(
  data=class_weight,
  orient='index',
  columns=['class_weight']
  )
df_class_weight.index = config.labels

library(reticulate)
py$df_class_weight

Table 4: Class weight

              class_weight
toxic          0.095900590
severe_toxic   0.009928468
obscene        0.052757858
threat         0.003061800
insult         0.049132042
identity_hate  0.008710911

It is also useful to define the steps per epoch for train and validation dataset. This step is required to avoid to not consume entirely the dataset during the fit, which happened to me.

steps_per_epoch = config.train_samples // config.batch_size
validation_steps = config.val_samples // config.batch_size

5.4 Fit

The fit has been done on Kaggle to levarage the GPU. Some considerations about the model:

• .repeat() ensure the model sees all the dataset.
• epocs is set to 100.
• validation_data has the same repeat.
• callbacks are the one defined before.
• class_weight ensure the model is trained using the frequency of each class, because our dataset is imbalanced.
• steps_per_epoch and validation_steps depend on the use of repeat.

history = model.fit(
  processed_train_ds.repeat(),
  epochs=config.epochs,
  validation_data=processed_val_ds.repeat(),
  callbacks=callbacks,
  class_weight=class_weight,
  steps_per_epoch=steps_per_epoch,
  validation_steps=validation_steps
  )

Now we can import the model and the history trained on Kaggle.

model = load_model(filepath=config.model)
history = pd.read_excel(config.history)

5.5 Evaluate

validation = model.evaluate(
  processed_val_ds.repeat(),
  steps=validation_steps, # 748
  verbose=0
  )

val_metrics <- tibble(
  metric = c("loss", "precision", "recall", "auc", "f1_score"),
  value = py$validation
  )
val_metrics

Table 5: Model validation metric

# A tibble: 5 × 2
  metric     value
  <chr>      <dbl>
1 loss      0.0542
2 precision 0.789 
3 recall    0.671 
4 auc       0.957 
5 f1_score  0.0293

5.6 Predict

For the prediction, the model does not need to repeat the dataset, because it has already been trained on all of the train data. Now it has just to consume the new data to make the prediction.

predictions = model.predict(processed_test_ds, verbose=0)

5.7 Confusion Matrix

The best way to assess the performance of a multi label classification is using a confusion matrix. Sklearn has a specific function to create a multi label classification matrix to handle the fact that there could be multiple labels for one prediction.

5.7.1 Grid Search Cross Validation for best threshold

Grid Search CV is a technique for fine-tuning hyperparameter of a ML model. It systematically search through a set of hyperparamenter values to find the combination which led to the best model performance. In this case, I am using a KFold Cross Validation is a resempling technique to split the data into k consecutive folds. Each fold is used once as a validation while the k - 1 remaining folds are the training set. See the documentation for more information.

The model is trained to optimize the recall. The decision was made because the cost of missing a True Positive is greater than a False Positive. In this case, missing a injurious observation is worst than classifying a clean one as bad.

5.7.2 Confidence threshold and Precision-Recall trade off

Whilst the KFold GDCV technique is usefull to test multiple hyperparameter, it is important to understand the problem we are facing. A multi label deep learning classifier outputs a vector of per-class probabilities. These need to be converted to a binary vector using a confidence threshold.

• The higher the threshold, the less classes the model predicts, increasing model confidence [higher Precision] and increasing missed classes [lower Recall].
• The lower the threshold, the more classes the model predicts, decreasing model confidence [lower Precision] and decreasing missed classes [higher Recall].

Threshold selection mean we have to decide which metric to prioritize, based on the problem we are facing and the relative cost of misduging. We can consider the toxic comment filtering a problem similiar to cancer diagnostic. It is better to predict cancer in people who do not have it [False Positive] and perform further analysis than do not predict cancer when the patient has the disease [False Negative].

I decide to train the model on the F1 score to have a balanced model in both precision and recall and leave to the threshold selection to increase the recall performance.

Moreover, the model has been trained on the macro avarage F1 score, which is a single performance indicator obtained by the mean of the Precision and Recall scores of individual classses.

\[ F1\ macro\ avg = \frac{\sum_{i=1}^{n} F1_i}{n} \]

It is useful with imbalanced classes, because it weights each classes equally. It is not influenced by the number of samples of each classes. This is sette both in the config.metrics and find_optimal_threshold_cv.

f1_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_f1, best_score_f1 = config.find_optimal_threshold_cv(ytrue, y_pred_proba, f1_score)

print(f"Optimal threshold: {optimal_threshold_f1}")

Optimal threshold: 0.15000000000000002

print(f"Best score: {best_score_f1}")

Best score: 0.4788653077945807

# Use the optimal threshold to make predictions
final_predictions_f1 = (y_pred_proba >= optimal_threshold_f1).astype(int)

Optimal threshold f1 score: 0.15. Best score: 0.4788653.

recall_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_recall, best_score_recall = config.find_optimal_threshold_cv(ytrue, y_pred_proba, recall_score)

# Use the optimal threshold to make predictions
final_predictions_recall = (y_pred_proba >= optimal_threshold_recall).astype(int)

Optimal threshold recall: 0.05. Best score: 0.8095814.

roc_auc_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_roc, best_score_roc = config.find_optimal_threshold_cv(ytrue, y_pred_proba, roc_auc_score)

print(f"Optimal threshold: {optimal_threshold_roc}")

Optimal threshold: 0.05

print(f"Best score: {best_score_roc}")

Best score: 0.8809499649742268

# Use the optimal threshold to make predictions
final_predictions_roc = (y_pred_proba >= optimal_threshold_roc).astype(int)

Optimal threshold roc: 0.05. Best score: 0.88095.

5.7.3 Confusion Matrix Plot

# convert probability predictions to predictions
ypred = predictions >=  optimal_threshold_recall # .05
ypred = ypred.astype(int)

# create a plot with 3 by 2 subplots
fig, axes = plt.subplots(3, 2, figsize=(15, 15))
axes = axes.flatten()
mcm = multilabel_confusion_matrix(ytrue, ypred)
# plot the confusion matrices for each label
for i, (cm, label) in enumerate(zip(mcm, config.labels)):
    disp = ConfusionMatrixDisplay(confusion_matrix=cm)
    disp.plot(ax=axes[i], colorbar=False)
    axes[i].set_title(f"Confusion matrix for label: {label}")
plt.tight_layout()
plt.show()

Figure 4: Multi Label Confusion matrix

5.8 Classification Report

cr = classification_report(
  ytrue,
  ypred,
  target_names=config.labels,
  digits=4,
  output_dict=True
  )
df_cr = pd.DataFrame.from_dict(cr).reset_index()

library(reticulate)
df_cr <- py$df_cr %>% dplyr::rename(names = index)
cols <- df_cr %>% colnames()
df_cr %>% 
  pivot_longer(
    cols = -names,
    names_to = "metrics",
    values_to = "values"
  ) %>% 
  pivot_wider(
    names_from = names,
    values_from = values
  )

Table 6: Classification report

# A tibble: 10 × 5
   metrics       precision recall `f1-score` support
   <chr>             <dbl>  <dbl>      <dbl>   <dbl>
 1 toxic            0.552  0.890      0.682     2262
 2 severe_toxic     0.236  0.917      0.375      240
 3 obscene          0.550  0.936      0.692     1263
 4 threat           0.0366 0.493      0.0681      69
 5 insult           0.471  0.915      0.622     1170
 6 identity_hate    0.116  0.720      0.200      207
 7 micro avg        0.416  0.896      0.569     5211
 8 macro avg        0.327  0.812      0.440     5211
 9 weighted avg     0.495  0.896      0.629     5211
10 samples avg      0.0502 0.0848     0.0597    5211

6 Conclusions

The BiLSTM model is optimized to have an high recall is performing good enough to make predictions for each label. Considering the low support for the threat label, the performance is not bad. See Table 2 and Figure 1: the threat label is only 0.27 % of the observations. The model has been optimized for recall because the cost of not identifying a injurious comment as such is higher than the cost of considering a clean comment as injurious.

Possibile improvements could be to increase the number of observations, expecially for the threat one. In general there are too many clean comments. This could be avoided doing an undersampling of the clean comment, which I explicitly avoided to check the performance on the BiLSTM with an imbalanced dataset, leveraging the class weight method.